Dynamic Models of Portfolio Credit Risk: A Simplified Approach

JOHN HULL AND ALAN WHITE

JOHN HULL is 3 professor of finance at Joseph L. Rotman School of Management, University of TortJiHo in ON. C.anada.
hull @ rotman.utoronto.ca
ALAN WHITE

is a professor of finance at Joseph L. i*,otman School of Management, University of Toronto in ON, C^anada.
awhite@rotman.utoronto.ca

We propose a simple dynamic model that is au attractive alternative to the (static) Gaussian copula model TIK model assumes that the hazard rate of a company has a deterministic drift with periodic impulses. The impulse size plays a similar role to default correlation in the Gaussian copula model. The model is analytically tractable and can he represented as a binomial tree. It can be calibrated so thai it exactly matches the term structure of GDS spreads and provides a good fit to CDO quotes of alt maturities. Empirical research shows that as the default eni'ironment worsens default correlation increases. Consistent with this research we find that in order to jit market data it is necessary to assume that as the default environment wor.':ens impulse size increases. We present both a hoino}^eneous and heterogeneous version of the model and provide results on the use of the calibrated mode! to value fonmrd CDOs, CDO options, and leveraged super senior transactions.

T
SUMMER 2U0S

he credit derivatives market has experienced meteoric growth since 1998. The most popular instruments are credit default swaps. These provide a payoff when a particular company defaults. However, in recent years portfolio credit derivatives have been attracting a lot of attention. These provide protection against the defaults experienced by a portfolio of companies. Statistics published by the Bank for International Settlements show that the outstanding notional principal for portfolio credit derivatives has grown from about $1.3 trillion

in December 2004 (20% of the notional principal for all credit derivatives) to about $10.0 trillion in December 2006 (35% of the notional principal for all credit derivatives). The most popular portfolio credit derivative is a collateralized debt obligation (CDO), in which a portfolio of obligors is defmed and a number ot tranches are specified. Each tranche is responsible for losses between U.% and U^% of the total principal for some U. and Uy As the market has developed, standard portfolios and standard tranches have been specified to facilitate trading. One example is the iTraxx portfolio, a portfolio ofl 25 investment-grade European companies with the notional principal (size of the credit exposure) being the same for each company. The equity tranche is responsible for losses in the range of 0 to 3% of the total notional principal. The mezzanine tranche is responsible for losses in the range of 3% to 6% of the total notional principal. Other tranches are responsible for losses in the ranges 6-9%, 9-12%, 12-22%, and 22100% of total principal. The buyer of protection pays a predetermined annual premium (known as a spread) on the outstanding tranche principal and is compensated for losses that are in the relevant range. (In the case of the equity tranche, the arrangement is slighdy different: The buyer of protection pays a certain percentage of the tranche principal upfront and then 500 bps on the outstanding tranche principal per year.)

OF DERLVAllVtS

Several other standard portfolios and associated tranches have been defined. For example, CDX NA IG is a portfolio of 125 investment-grade North American credit exposures. (The tranches for this portfolio are 0 to 3%, 3 to 7%, 7 to 10%, 10 to 15%, 15 to 30%, and 30 to 100%) The most popular life of a CDO is 5 years. However, 7-year, 10-year, and to a lesser extent, 3-year CDOs now trade fairly actively. Default correlation is critical to the valuation of portfolio credit derivatives. Moody's statistics show that between 1970 and 2005 the default rate per year ranged from a low of 0.09% m 1979 to a high of 3.81% in 2001. This tendency of defaults to cluster has been studied by a number of researchers. One possible explanation is that default rates of all companies are influenced by one or more macroeconomic factors. Another is that defaults are "contagious" in the sense that a default by one company may induce other corporate failures. Das, DufFie, Kapadia, and Saita [2007] argue that contagion accounts for some part ofthe default clustering that is observed in practice. The standard market model for valuing portfolio credit derivatives assumes a simple one-factor model for a company's time to default. This is referred to as the Gaussian copula model. Its origins can be found in Vasicek [1987], Li [2000], and Laurent and Gregory [2005]. The Gaussian copula model is a static model. A single normally distributed variable determines the default environment for the whole life of the model. When the variable has a low value, the probability of each company defaulting during the life ofthe model is relatively high; when it has a high value, the probability of each company defaulting is relatively low. The model does not describe how the default environment evolves. Many alternatives to the Gaussian copula such as the ^-copula, the double-r copula, the Clayton copula, the Archimedean copula, the Marshall-Olkin copula, and the implied copula have been suggested. In some cases these models provide a much better fit to market data than the Gaussian copula model, but they are still static models. The availabihty of CDO data for multiple time horizons presents researchers with an interesting and important challenge; develop a dynamic model that fits market data and tracks the evolution of the credit risk of a portfolio. Dynamic models are important for the valuation of some structures. For example, American options on tranches of CDOs cannot be valued without a dynamic model. There are two types of dynamic credit risk models, which we will refer to as "specific" and "general" models.
10

In a specific model the company or companies being modeled remain the same through time. In a general model they do not remain the same but are defined to have certain properties. A model ofthe evolution ofthe credit spread for a particular company or the evolution of losses on a particular portfolio is a specific model. A model of the evolution of the average credit spread for A-rated companies or ofthe mezzanine spread for CDX NA IG is a general model. This article is concerned with the development of a specific dynamic model for portfolios. Extensions ofthe Merton [1974] structural model provide one approach for developing a specific dynamic model. Correlated processes for the values ofthe assets of the underlying companies are specified and a company defaults when the value of its assets reaches a barrier. The most basic version of the structural model is very similar to the Gaussian copula model. Extensions ofthe basic model have been proposed by Albanese, Chen, Dalessandro, and Vidler [2005], Baxter [2006], and Hull, Predescu, and White [2005]. Structural models have the advantage that they have sound economic underpinnings. Their main disadvantage is that they are difficult to calibrate to market prices and usuaUy have to be implemented with Monte Carlo simulation. Reduced-form models provide an alternative to structural models. The most natural reduced-form approach to developing a dynamic model is to specify correlated diffusion processes for the hazard rates ofthe underlying companies. Our own experience and that of other researchers is that it is not possible to fit market data with this type ot model. This is because there is a limit to how high the correlation between times to default can become. This has led researchers to include jumps in the processes for hazard rates. DufFie and Garleanu [2001], for example, assume that the hazard rate of a company is the sum of an idiosyncratic component, a component common to all companies, and a component common to all companies in the same sector. Each component follows a process with both a diffusion and a jump component. Other reduced-form approaches are provided by Chapovsky, Rennie, and Tavares [2006], Graziano and Rogers [2005], Hurd and Kuznetsov [2005J, and Joshi and Stacey [2006]. Another approach to developing dynamic models involves the development of a model for the evolution of the losses on a portfoho. This is sometimes referred to as the "top down" approach. The behavior of individual companies in the portfolio is not considered. Sidenius,

DYNAMIC MODELS OF PORTFOLIO CREDIT RISK: A SIMPLIFIED APPROACH

SUMMER 2008

Piterbarg, and Andersen |2004J use concepts from the Heath, Jarrow, and Morton [1992] interest rate model to suggest a complex general no-arbitrage approach to modeling the probability that the loss at a future time will be less than some level. Bennani [20051 proposes a model of the instantaneous loss as a percentage of the remaining principal. Schonbucher [2005] models the evolution ofthe loss distribution as a Markov chain. Errais, Gieseke, and Goldberg [2006| suggest a model where the arrival rate of defaults experiences a jump when a default happens. In LongstaiTand Rajan [2006], the loss follows a jump process where there are three types ofjumps: firm-specific, industry, and economy-wide. Putyatin, Prieul, and Maslova [2005] suggest a model where the mechanism generating multiple defaults resembles the kinetics of certain chemical reactions. Walker [2007] uses a dynamic discrete-time multi-step Markov loss model. Our objective in this article is to develop a model that is easy to implement and easy to calibrate to market data. The model is developed as a reduced-form model but can also be formulated as a top-down model. Under the model the hazard rate for a company follows a deterministic process that is subject to periodic impulses. This leads to a jump process for the cumulative hazard rate (or equivalently for rhe logarithm ofthe survival probability). CDOs, forward COOs, and options on CDOs can easily be valued analytically using the model. For other instruments a binomial representation ofthe model can be used. We propose a calibration method where credit default swap spreads are exactly matched and CDO tranche quotes are matched as closely as possible. THE MODEL For any particular realization ofthe hazard rate, li{t), between time zero and time i we can define

The variable S{t) is the cumulative probability of survival by time / conditional on a particular hazard rate path between time 0 and time I. A common approach to building a reduced-form model is to define a process for h{t) for each company over the life ofthe model. We choose instead to defme a process for S{t) for each company. The process for S{t) provides the same information as the process for h{t). However, it is much easier to work

with because it leads to a much more straightforward way of valuing CDOs. (As we wiU see, the process for h{t) that corresponds to the process for S{t) that we assume is one that has a deterministic drift and periodic impulses.) The default probability between time zero and time /, as seen at time zero, is the expected value of 1 - S{t). It is important to specify what is known in this model. The underlying state variable is S{t). While S{t) is not directly observable it may be inferred from the prices of credit-sensitive contracts. We also know how many defiuilts have occurred by time f. In developing and applying the model, we will usually assume that at any given time, f, we know both S{t) and the total number of defaults up to time t. This defmes the filtration.' Note that we do not assume that we know which particular companies have defaulted. In the homogeneous version of the model this additional information would ofcourse be irrelevant. Unfortunately, the heterogeneous version of the model becomes computationally intractable if an attempt is made to incorporate this additional information. The model assumes that most ofthe time the default probabilities of companies are predictable and defaults are independent of one another. Periodically there are economy-w^ide shocks to the default environment. When a shock occurs each company has a non-zero probability of default. As a result there are liable to be one or more defaults at that time. It is these shocks and their size that create the default correlation. The model is ofcourse a simplification of reality. In practice shocks to the credit environment do not cause several companies to default at exactly the same time. The defaults arising from the shocks are usually spread over several months. However, the model's assumption is reasonable because it is the total number of defaults rather than their precise timing that is important in the valuation of most portfolio credit derivatives. Another siniphfication is that shocks to the credit environment affect all companies. In practice they are likely to affect just a subset of companies in the portfolio. However, as an approximation we can think of a shock afTecting a subset of companies as being equivalent to a smaller shock affecting all companies as far as its effect on the number of defaults is concerned. For ease of exposition in explaining the model we assume homogeneity so that all companies have the same default probabilities. We also assume that the recovery rate is constant. Later we explain how these assumptions can he relaxed.

SUMMER 2008

THL joukNAL (.)F DERIVATIVES

11

Consider a portfolio of obligors with total notional principal L. The protection seller for a tranche of a CDO provides protection against losses on the portfolio that are in the range a^ L to ii^L for the life of the instrument. The parameter d^ is known as the attachment point and the parameter a^ is known as the detachment point. The protection buyer pays a certain number of basis points on the outstanding notional principal of the tranche. This principal equals a^L aj^ initially and declines as losses in the range a^ to a^ are experienced. As explained in Hull and White |2006|, the key to valuing a CDO tranche at time zero is the calculation of the expected tranche principal on payment dates. The expected payment by the buyer of protection on the CDO tranche on a payment date equals the expected tranche principal on the payment date multiplied by the spread. The expected payoffby the seller of protection between two payment dates equals the reduction in the expected tranche principal between those dates.- The expected accrual payments required in the event of a default between two payment dates can be calculated from the reduction in the expected tranche principal between the dates and an assumption about when the reduction occurs. As we will see the expected tranche principal can be easily calculated from S{t). This is why a model ofthe behavior of 5{?) is much easier to work with than a model ofthe behavior of b{t). The properties of S{t) are that 5{0) - 1, S is nonincreasing in time, and S{t) > 0 for all f. A convenient transformation of S is to define a variable X = ln(S) that has the properties that X(0) = 0 and X is non-decreasing in time with no upper bound. We assume that in a riskneutral world X follows a jump process with intensity X and jump size H? dX = ^dt + dq (1)

The term dl is an impulse that has intensity X. The impulse takes the form of a Dirac delta function. The effect of an impulse at time t is to cause the hazard rate to become infinite in such a way that the integral of the hazard rate over any short interval around time f is finite. The jumps in the model can lead to several companies defaulting at the same time. For example, suppose that S decreases from 1 to 1 g as a result of a jump at time t. There was no chance of default before the jump, but each company has a probability q of defaulting at time /. If there are N companies in the portfolio the probability that n of them will default at time f is then

N!

Another model which allows multiple defaults at the same time is the generalized Poisson loss model of Brigo, Pallavicini, and Torresetti [2007]. In this model there are several independent Poisson processes. An integer z. is associated with the (th process. When an event occurs in the ith process, z. defaults occur simultaneously. The model provides a good fit to data on CDOs with several maturities but does not have the dynamic structure of our model.

Notation
We will use the following notation: fp t^ T h e probability of exactly J j u m p s between times (j and ^2 ('2 -^ ^j) The probability of exacdyj jumps between time zero and time t (= p{J, 0, f)) H, The size of the J t h j u m p in X fp (,) ,0 T h e probability of exactly ti defaults between times /, and t-y T h e probability of exactly defaults between, time zero and time t {=<^{i!, 0, t)) The cumulative probability of survival by time t when there have beenJjumps between time zero and time /

P{J, f)

In any short interval of time, Af, dq H with probability AAr, and dq = 0 with probability 1 - AAt. The nondecreasing nature of X requires that /i > 0, and H > 0. We will assume that // and X are functions only of time and H is a function only ofthe number of jumps so far/ For the purpose of using the model we do not need to explicitly consider the hazard rate process. However it is interesting to note that the process followed by the hazard rate, h(t), is more extreme than the jump process considered by other researchers. It is dl

S{J, t)

12

DYNAMtc MODELS OF PORTFOLIO CREHIT KISK; A

APPROACH

SUMMER 2008

A{t)
(4)

M{t)

Expected principal on a tranche at time t when the initial tranche principal is SI Present value of $1 received at time t

The probability of if defaults in the portfolio by time t conditional onjjunips is

O(, t\f) = /.(, N, 1 - SiJ, t))
(5)

t'> 0

Tlie remaining CDO tranche principal at time t when there have been n defaults. The initial tranche principal is $1

where b is the binomial probabihty function; b{n, N, q) =

N\

Valuation of a CDO CDOs can be valued analytically using the model. Since the payments and payoffs associated with a CDO do not depend on decisions by the buyers and sellers of protection, the value of the CDO is the same under all filtrations. As before, let a^ andrt,^be the tranche attachment and detachment points, respectively. Defme
a,N

n\{N~ny.

The expected principal on the CDO tranche at time t conditional onjjunips is

(6)

The unconditional expected principal at time t is therefore (7) Let the payment times be f,, ^, ... , r , define t,, - 0, and
r J I' 2' ' III' 11 '

\-R

"

\-R

=-^

where Af is the number of companies in the portfolio and R is the fixed recovery rate. Denote the smallest integer greater than x by g{x). For a tranche with an initial principal of $1 the tranche principal at time / when there have been n defaults is
when {ii, t) =
H 1.

assume that defaults always happen halfway through the period between payments. If the initial principal is SI the present value of the regular payments that are made on payment dates is sA where s is the spread and (8)

"<^("j.) ^ ( " L ) - " *^ S^i"ii. II > uln..)

when

The present value of the accrual payments made in the event of a default is sB where
(2)
(9)

when

1 he probability ofj jumps between time zero and time f is where /^ - 0.5(/(, + ( j). The present value of the pay^ offs arising from defaults is
(10)

(3)

The value of 5 at time / if there have beenjjumps is

SUMMKR 20i)H

THE JOURNAL OF DERIVATIVES

13

The total value ofthe contract to the seller of protection is i.^ + sB- C. The breakeven spread is C/{A + B).

i\,{n,t\J)P{J,t)
n) =

(14)

The Loss Distribution

The model has been presented as a reduced-form model. However, it can be converted to a top-dov^'n model where the process for the loss is modeled. Because we are assuming a constant recovery rate it is sufFicient to model the number of defaults n. In this version ofthe model the filtration is different from that given at the beginning of this section. It is assumed that at time t we know the number of defiiults but v^^e do not knovt' survival probabilities. The process for survival probabilities becomes nothing more than a convenient tool for generating the model. The proportion ofthe original portfolio lost by time t when the number of defaults is n is Equations (12) and (13) can be used to calculate the probability that there will be n^ defaults at time t^ conditional on /(| defaults and_/, jumps by time /,:

Using Equation (14) we obtain the transitio)] probability from H, defaults at time r, to n, defaults at time (16)

(1-R) N
First consider the unconditional distribution for at time /. This can be calculated from Equations (3) and (5)

This equation defmes the process for the number of defaults (or equivalently the loss) for the portfolio. ALTERNATIVE VERSIONS OE THE MODEL A N D THEIR CALIBRATION In this section we first present a version ofthe model where (similar to Black-Scholes) there is just one free parameter. We then discuss extensions to this simple model and present a three-parameter version ofthe mode! that is designed to provide a good fit to all C D O spreads ot all maturities. The calibration ofthe model to the market will be illustrated using the data in Exhibit 1 for iTraxx and CDX NA IG on January 30, 2007.^ We assume a recovery rate of 40%. The term structure of CDS spreads is assumed to be a piece-wise linear function. It equals the three year spread for maturities up to three years, between years three and five the spread is interpolated between the threeand five-year CDS spreads, and so on. Payments on CDOs and CDSs are assumed to be quarterly in arrears.

(11) This is a mixture of Poisson distributions. The loss distribution given by the model has thin tails when jumps are small and fat tails when they are large. This means that the model is a flexible tool for handHng a variety of default correlation environments. The probability off; defaults between times t^ and ^2 conditional on Jj jumps by time Z^,/., jumps by time t^, and ttj detaults between times zero and ij is

(12) The probability ofjjumps between times t. and t^ is

Zero Drift, Constant Jumps, Time-Dependent Intensity

MJ^V^) =

(13)

and from Bayes' rule the probability of J jumps by time t conditional on /) defaults by time t is

A particularly simple version of the model is the case in which fJ{t) = 0 and the jump size is constant. For any given value of the jump size, M, the jump intensity X(t) is chosen to match the term structure of CDS spreads.

14

DYNAMIC MODELS OF PORTFOLIO CREDIT RISK: A SIME'LIFIED APPROACH

SUMMER 2O()S

EXHIBIT 1 Market Quotes for Standard Tranches of iTraxx and CDX N A IG, January 30, 2007
iTraxx CDO tranche quotes January 30, 2007. 3yr 0 0.03 0.06 0.09 0.12 0.03 0.06 0.09 0.12 0.22 Index n/a n/a n/a n/a n/a 15.00 5yr 10.25 42.00 12.00 5.50 2.00 23.00 7yr 24.25 106.00 31.50 14.50 5.00 31.00 10 yr 39.30 316.00 82.00 38.25 13.75 42.00

The imphedjump size is a measure of default correlation. As the jump size approaches zero the default correlation approaches zero. As the jump size becomes large the default correlation approaches one. Exhibit 3 compares the implied jump sizes reported in Exhibit 2 with the tranche (or compound) correlations implied from the

CDX NA IG CDO tranche quotes January 30, 2007. 3yr 0 0.03 0.07 0.10 0.15 0.03 0.07 0.10 0.15 0.30 Index n/a n/a n/a n/a n/a 19.00 5yr 19.63 63.00 12.00 4.50 2.00 31.00 7yr 38.28 172.25 33.75 14.50 6.00 43.00 10 yr 50.53 427.00 96.00 43.25 13.75 56.00

EXHIBIT 3 Implied Jump Size Using the One-Parameter Model Compared with the Implied Tranche Correlation from the Gaussian Copula Model for 5-, 7-, and 10-Year Tranches of iTraxx, January 30, 2007
S-Year Quotas -- 1 mplied Correlalion -e- Implied Jump

We assume that the value of A{f) is constant between CDO/CDS payment dates and work forward in time choosing As so that the CDS term structure is matched. This version ofthe model has just one free parameter: the jump size. The jump size can be implied from a CDO market quote or vice versa. The procedure for calculating the value of a CDO tranche was described earlier. The jump sizes implied from the iTraxx market quotes in Exhibit 1 are shown in Exhibit 2. The numbers in this table are analogous to the numbers in a volatility surface that is determined from market quotes for option prices using the Black-Scholes formula. Just as option traders monitor volatility surfaces, credit derivatives traders can monitor the jump size surface. EXHIBIT 2 Implied Jump Sizes for iTraxx on January 30, 2007 for the One-Parameter Model where the Drift is Zero and the Jump Size is Constant
3yr
0 003 0.06 0.09 0.12 0.03 0.06 0.09 0.12 0.22 5 yr 0.0247 0.0120 0.0336 0.0578 0.0981 7yr 0.0221 0.0054 0.0268 0.0501 0.0900 lOyr 0.0221 0.2378 0.0112 0.0340 0.0749

610 9 Tranche 7-Year Quoles 0.25

9 to 12

12 lo 22

3106

61o9 Tranche 10-Year Quotes

0.45 0.40 0.35 0.30 0.25 0.20 0.15 0.10 0.05 0,00 6 to 9 Tranche 9 to 12 12lo22 ' Implied Correlation Implied Jump

n/a
n/a n/a

T^TVV

n/a
n/a

SUMMER 201)8

THE JOURNAL OF DERIVATIVES

15

tranche quotes using a Gaussian copula model. It can be seen that the two exhibit very similar patterns. Results are similar for CDX NA IG. The advantage of calculating an implied jump size rather than an implied copula correlation is that the jump size is associated with a dynamic model whereas the copula correlation is associated with a static model. Extensions of the Constant Jump Model There are a number of ways the model we have just considered can be extended. The most obvious extension is to allow a non-zero drift. In this case the jump size and intensity can be chosen to fit CDO quotes while the drift is selected to fit the CDS term structure. In practice this does not provide material improvement over the zero-drift case. We can obtain a better fit to CDO prices for a particular maturity by having multiple jump processes each with its own size and each with its own intensity or by having a single jump process in which the jump size is a random variable. This would be in the spirit of Longstaff and Rajan [2006] or Brigo et al |2U06|. These approaches work reasonably well in fitting panels of CDO quotes but pose computational problems in the convolution of the survival probabilities. When either of these approaches is implemented, the best fit to market CDO quotes arises when there is a high probability of small jump sizes and a low probability of large jump sizes. These approaches do not have the property that large jump sizes are associated with adverse states of the world. Rather than developing the model along these lines we have chosen to consider a version of the model that is particularly easy to unplement and involves relatively few parameters. A Three-Parameter Model Empirical evidence suggests that default correlations are stochastic and increase in adverse credit conditions. For example, Servigny and Renault [2002J, who look at historical data on defaults and ratings transitions to estimate default correlations, find that the correlations are higher in recessions than in expansion periods. Das et al. [2006] employ a reduced-form approach and compute the correlation between hazard rates. They conclude that correlations increase when hazard rates are high. Hull et al. [2005] show that a structural model fits

EXHIBIT 4 Model Calibration Results

Errors resulting from calibration of three-parameler model to the iTraxx data in Exhibit 1 for January 30, 2007, (For example, the quote for the 3% to 6% 10-year tranche is 316 and the best fit spread is 314.63.)

a
0 0.03 0,06 0.09 0,12 0.03 0.06 0.09 0.12 0,22 Index

3yr n/a n/a n/a

5yr 1.34 0.37 -0.54 -1,01 -0,47 0.00

7yr
2.75 3,12 -2.69 -1.55 -0.21 0.00

10 yr 4.32 -1.37 -1.92 -0.12 1.28 0.00

n/a
n/a 0.00

Errors resulting from calibration of the three-parameter model to the CDX NA IG data in Exhibit 1 for January 30, 2007.

a.
0 0.03 0,07 0.10 0,15

a
0.03 0.07 0.10 0.15 0.30 Index

3yr n/a n/a n/a n/a n/a 0.00

5yr 1.63 -4,01 2.30 4.00 0.69 0.00

7yr 3.20 -2,16 3.39 479 1.28 0.00

10 yr 2.85 1.99 2.51 1,44 5,55 0.00

CDO data well when asset correlations are positively related to the default rate. Their work is consistent w^ith that of Ang and Chen [2002], who find that the correlation between equity returns is higher during a market downturn. This research suggests that a model where the jump sizes in X are larger in adverse market conditions might fit market data better than a constant jump size model. To test this we relax the constant jump size assumption and explicitly build in the property that large jump sizes (high correlation) are associated with low survival probability states. The size of the Jth jump is given by H, = where H,, and are positive constants. The intensity of the process, X, is a constant and the drift, //(/), is determined to match CDS spreads. In calibrating the model the objective is to fmd values of Hjj, y9, and A that minimize the sum of squared differences between market tranche spreads and model tranche spreads. The procedure involves repeatedly 1) choosing trial values of H,j, p, and X, 2) calculating the function so that the term structure of index spreads

16

l>i'NAMi(: MODELS OF PORTFOLIO CREDIT RISK: A SIMI'LIF-IKD

SUMMER 201 )S

is matched for the trial parameters, and 3) calculating the sum ot squared differences between model spreads and market spreads for all tranches of all maturities (15 spreads in total). An iterative procedure is used to find the values of H||, p, and X that lead to the sum of squared differences being minimized when this three-step procedure is used. For the iTraxx data in Exhibit 1 the best fit parameter values are H,, = 0.00223, P= 0.9329, and >l= 0.1486. The corresponding values for the CDX NA IG data are H^. = 0.00147. P = 1.2813. and X = 0.1310. The pricing errors arc shown in Exhibit 4. The model fits market data well^ much better than versions of the model where the jump size is constant. Exhibit 5 shows the loss distribution 3, 5, 7, and 10 years in the future for iTraxx as seen at time zero based on the calibrated model. As the time horizon increases, che probability mass of the distribution shifts to the right and becomes more spread out. There is also some fine

structure in the distribution. To make this visible, all probabilities for losses greater than 9% are scaled up by a factor of 100. This reveals a complex shape to the right tail of the loss distribution. Results for CDX NA IG are similar. The values of H. are initially fairly small but increase fast. For example in the case of iTraxx on January 30, 2007, H3 = 0.057, H^ = 0.386, and H^ = 2.513. There is a small probability of low values of S being reached. For example, the probability that S is less than 0.90 at the end of 5 years is about 0.007, and the probability that it is less than 0.75 is about 0.0001. Some of these values inay seem extreme. However, they are consistent with tbe results in papers such as Hull and White [2006], which show that it is necessary to assign a very low, but non-zero, probability to a very high hazard rate in a static model in order to fit market quotes. !t should be recalled that the results shown here are risk-neutral probabilities that are inferred from tranche

EXHIBIT 5 Unconditional Loss Distribution for iTraxx on January 30, 2007, at Four Maturities

Unconditional Loss Distribution at 4 Maturities

3-Years - B - 5-Years -^^7-Years -*-10-Years
0.30

0,20 ^

Probabilities for iosses greater than 9% multiplied by 100,

I
\ / * * \

0,10-

4
0,00 * 0% 3% 6% 9%
, ,nnn nnnrii ,g9^ri n i LJ U L

12%

15%

18%

21%

24%

Loss {% of Total Notional)

Now: Results arc hased on the ihree-paramcu-r model calibrated to the market data in Exhihit 1.

SUMMER 2008

THE JOURNAL OF DERIVATIVES

17

prices. Consider the 2 bp cost of protection for the 5-year 12 to 22% tranche of iTraxx. Given the recovery rate that is assumed, the buyer of protection only receives payoffs wben more than 20% ofthe entire portiblio has defaulted and payments for protection only stop when the tranche is wiped out after 36.7% of the portfolio has defaulted. Suppose there are only two possible outcomes: the tranche is wiped out with probability/; or is untouched with probabihty 1 -p. If tbe tranche is wiped out, the buyer of protection receives a payoff of $1; if the tranche is untouched, he pays 2 bps per year for 5 years on a notional of }EI. Ignoring discounting, the value of/J that makes this a fair contract is about 0.001, which is roughly consistent with the probabilities implied by the calibrated model. This calibration procedure was repeated for all the iTraxx tranclie data that was avaikble from Reuters between July 4, 2006, and January 10, 2007. This data includes the spreads on 5-, 7- and 10-year C D O tranches as well as 3to l()-year index spreads. All 15 C D O tranche spreads were available on 51% ofthe days, 14 spreads were available on 34% ofthe days, 13 spreads were available on 13% ofthe days, and on 2% ofthe days only 12 spreads were quoted. Tbe quality of the fit on each day was similar to that reported for the January 30, 2007, market data. Exhibit 6 shows lO-day moving averages for the parameter values.'' APPLICATIONS OF THE MODEL In this section we show how the three-parameter model we calibrated to iTraxx market data can be used to value a number ot different instruments.

based on a portfolio that will be selected to be investment-grade at the start ofthe protection period covered by the forward contract. Forward contracts can be valued analytically using the model in a similar way to CDOs. As with the CDO, since the payments and payoffs associated with the forward contract do not depend on decisions by the buyers and sellers, the value ofthe forward contract is the same under all filtrations. Suppose that the forward contract lasts between payment times / and ( . Define/1(?), B(/), and C{tJ similarly to A, B, and C in Equations (8), (9), and (\0) except that they accumulate expected payoffs and payments between times t and / rather than between time zero and time ( :

The total value ofthe forward C D O contract to the seller of protection w^hen the tranche spread is s is sA{t ) + breakeven forward spread is The breakeven spreads for forward contracts on C D O tranches that mature in five years are shown in Exhibit 7. The tranche spreads rise fairly fast. One reason for this is that the CDS spreads are upward sloping so that default probabilities tend to increase as time passes. Another reason is that, in the case of all tranches except the equity tranche, there is very little probability of loss in the first one or two years. It follows that when these years are excluded, the spread for the remaining years increases. O p t i o n s on CDO Tranches A European option on a C D O tranche is an option to buy or sell protection for a particular tranche at a particular strike spread. The option expires at time f^ and if ^, exercised, the protection lasts between times t and f . As with the forward C D O contract, the underlying portfolio is defined at time zero and remains unchanged. Upon

Forward CDOs
Consider first forward CDOs."" As discussed earlier, the model in this article is a specific dynamic model. We are modeling defaults for a portfolio of companies that is defined at time zero and remains unchanged. The forward C D O spreads that we calculate using the model are the spreads for this portfolio. For example, the 3 x 2 forward spread for the mezzanine (3% to 6%) tranche is the spread that must be paid in years 4 and 5 on the remaining amount ofthe original tranche principal in order to provide protection. The protection is against those default losses that occur in years 4 and 5 and that are between 3% and 6% ofthe principal ofthe original underlying portfolio. The calculated spread is not that for a forward start de nouo iTraxx contract. The latter contract would be

18

DYNAMIC MODEL-I OI- PUKTFOLIO CREUIT RISK: A

APPROACH

SUMMER

EXHIBIT 6 Parameters H^, j5, and A Estimated for iTraxx Between July 4,2006, and January 10, 2007
Jump Parameters -Beta (left-hand scale) -e-HO (right-iiand scale)
1.5 0.004

1.2
0.003

0.002

0.6

0.001 0.3

0.0 4-JUI-06 3-Aug-06 2-Oct-06 1-Nov-06 1-Dec-06 31-Dec-06

0.000

Jump Intensity
0.20 0.16 0.16 0.14 0.12 0.10 0.08 0.06 0.04 0.02 0.00 4-Jul-06 3-Aug-06 2-Sep-06 2-OC1-06 1-Nov-06 1-Dec-06 31-Dec-06

Option exercise, the spread is applied to the remaining principal (ifany) ofthe tranche. Like forwards, European options can be valued analytically using the model. Since the buyer of the option must decide when to exercise the option, the information available (the filtration) will affect tbe value ofthe option. It will be recalled that

we are working in a filtration where both S{t) and the number ot defaults between time zero and time t are known at time t. This information will therefore be used to determine exercise decisions. The number of defaults is directly observable by the holder of an option, and we assume that the prices of CDSs and CDO tranches reflect

SUMMER. 2UU8

THE JOURNAL OF DERIVATIVES

19

EXHIBIT 7 Breakeven Tranche Spread for Forward Start CDO Tranches on iTraxx on January 30, 2007.
Tranche Start 1,0 2.0 3.0 4,0 4.5

Breakeven Tranche Spreads 0 0.03 0.06 0.09 0,12 Index 0.03 0,06 0.09 0.12 0.22 11,5 54.0 14.7 5.8 2.0 25.3 11.3 70.1 19,4 7.7 2.6 29.1 11.1 93,2 26,1 10.6 3,7 36.7 6.4 124.4 35.2 14.8 5.3 41.4
1.

3.4 144.2 40,7 17,5 6.3 43,7

Note: TItc tranches mature in five years. Results are based on the three-parameter model calibrated to the market data in Exhibit

the state ofthe economy, S{t). These prices are observable and can be used to formulate an exercise decision. An option to buy protection will be exercised at time t if the cost of protection on the tranche at time t^^ is higher than the strike spread. Given the structure of our model, knowing the value of S at time t is equivalent to knowing the numher of jumps / before time t . Conditional on n defaults and / jumps by time f^ the expected tranche principal at time ^, V (^' -^ ") '^" b^ calculated from Equations (2) and (15) as .

Suppose the strike spread for a European call option is Sf^. The present value ofthe option when there have been n defaults and J jumps by time ( is max[C{f \n J )-s,.A{t \ti J )-s,.B(t \n J ),01

The value of the call option can therefore be calculated from Equations (3), (4), and (5) as

t \fi , J
H I U' J U

\n ,J ) , 0 t

From this the present value of payments, accrual payments, and payoffs condition on n^ ^ndj^^ can be calculated:

Similarly the value of a European put option is

,t I / )P(/ ,r
II'

(H

In ,J
II I tl^ J \l

J )=

n ,J J

-',. ".='"'

I I I - > II '

^ J

II'

II

= 0.5

n,.

= I [^(V, I'V J,. U 20

> U]

When the strike spread is set equal to the forward spread, an at-the-nioney option is created and the put and call have the same price. Exhihit 8 shows the prices in basis points of at-the-nioney options with varying maturities on a five-year CDO tranche. For example, a one-year at-the-money option to buy at-the-money protection on the 6% to 9% tranche for the period from one year to five years costs 23.2 bps or 0,232% ofthe initial (time zero) tranche principal. The option strike

DYNAMtc M O D E L S OF P O R T F O L I O C R E D I T HJSJC: A SIMPLIFIEIJ AI'PR(.>ACH

S U M M E R 2(K)M

EXHIBIT 8 Prices in Basis Points of at-the-Money European Options on iTraxx CDO Tranches on January 30, 2007
Option Expiry in Years 1,0 0.03 0.06 0.09 0.12 0.06 0.09 0,12 0,22 67.8 23.2 9.7 3,7 2.0 91,3 29.8 12.2 4.4 3.0 89.7 30.7 13.3 5.0 4.0 68.3 23.1 10.0 3.8 4,5 41,4 13.5 6.1 2.4

Again the results are based on Hull and White |2()()7]. If the assumption of lognormality held exactly the implied volatility for an option would be the same for all strike prices. The exhibit shows that the variation of implied volatility with the strike price is quite small. This suggests that the lognormality assumption in HuU and White [20071 is (at least for January 30, 2007, data) approximately consistent with the model in this article. Leveraged Super Senior Transaction Our final application ofthe model is to a leveraged super senior (LSS) tranche with a loss trigger. This is a CDO tranche that is automatically cancelled when losses reach some level. On cancellation the tranche is marked to market and the seller of protection must pay the buyer an amount equal to the value of the tranche. However, the total amount paid by the seller (including any losses for which the seller is responsible prior to the cancellation date) is capped at a fi^action x ofthe tranche notional. Since the payments and payoifs associated with the LSS do not depend on decisions by the buyers and sellers the value ofthe contract is the same under all filtrations. Because we are assuming a constant recovery rate ot 40%, the loss that triggers cancellation can be translated to a number of defaults. We suppose that cancellation is triggered when the number of defaults reaches *. The parameters defining the LSS on a particular portfolio are therefore the tranche attachment point a., the tranche detachment pointrt,.^,the spread s, the maturity T, the loss cap X, and the number of losses triggering cancellation n'.'^ To simplify matters we only consider cases in which the number of defaults that trigger a termination. ;/', is less than 125(7^/(1 - R). This ensures that the seller is not responsible for any losses prior to the termination date.'" The tranche in a LSS is generally a senior one. The seller of protection considers it highly unlikely that the tranche will experience losses but does not find selling protection on the tranche to be appealing because the spread is small. The LSS provides a way of leveraging the spread. Suppose that the tranche principal is $10 million and the spread is 15 bps. Selling protection on the tranche in the usual way requires %\0 million to be deposited by the seller of protection at an interest rate of LI13OR, so that if the tranche does not experience defaults, tbe interest earned is LIBOR plus 15 bps. Consider an LSS with X - 0.1. Only $1 million need to be deposited at the

Now: TIte tranches muttire injiuc years. Results are based on tlic tlmr-parameler model calibrated to the market data in Exhibit I.

spreads are the corresponding forward spreads given in Exhibit 7 / Tranches with higher spreads (lower attachment points) have higher option values. The value ofthe option initially increases as we increase the option maturity and then starts to decline to reflect the diminishing maturity ofthe underlying. This is similar to what we observe for options on bonds and swaps. As discussed in Hull and White [20{)7|, if we are willing to assume that CDO tranche spreads are lognormally distributed, it is possible to derive an analytic expression for the prices of European put and call options on CDO tranches. Exhibit 9 uses the result in Hull and White |2()()7] to calculate tlie implied spread volatihties from the prices in Exhibit 8. Usually, the implied volatility increases with option maturity. Exhibit 10 shows the implied volatilities for twoyear options on five-year CDO tranches for strike prices between 75% and 125% ofthe forward tranche spread. EXHIBIT 9 Implied Volatilities of at-the-money European Style Options on iTraxx CDO Tranches on January 30, 2007
Option Expiry in Years a. 0.03 0.06 0.09 0,12 0.06 0,09 0.12 0.22 1.0 96.1% 123,2% 133.0% 149.3% 2.0 100,9% 122.6% 128.6% 137.3% 3.0 96.8% 125.6% 139,4% 160.5% 4.0 104,3% 137.2% 144,8% 160.6% 4.5 107.5% 135,3% 148.9% 181,8%

Note: The irarichef mature in /ice years. Results tire hased oti the threeparameter tnodel calibrated to lite market data in Exhibit 1.

SUMMER 2008

OF DERIVATIVES

21

EXHIBIT 10 Implied Volatilities for 2-Year European Options on iTraxx CDO Tranches for Strike Prices Between 75% and 125% of the Forward Spread on January 30,2007

on each payment date (i.e., for each k, T. f, ^ for some () In practice this is achieved by creating V equal time steps between each payment date for some integer V. Note that the binomial tree as it is presented in Exhibit 11 only reveals 5(0- As a CDO Tranche result, it would produce different values for options on a CDO tranche than those in 12 to 22% 6 to 9% 9 to 12% KIF 3 to 6% Exhibit 8." To value an option where the exercise decision depends on both 5(0 and the 125.6% 132,9% 143.5% 0.75 100.1% number of defaults to date, it would be neces125.2% 132,1% 142.3% sary to construct a binomial tree for S where 0.80 100.6% there are N + 1 states at each S-node corre124.7% 131,3% 141.1% 0.85 100.9% sponding to the N + 1 alternative values for the number of defaults at the node. The use 101.1% 124,1% 130,5% 139.9% 0.90 of this type of tree was proposed by Hull and 123.4% 129.5% 138.6% 0.95 101.0% White [1993]. It is now sometimes referred to as a "binomial forest." As one rolls back 100.9% 122.6% 128.6% 137.3% 1.00 through the tree, calculations are carrieci out for all N + 1 states at each node. 100.6% 121.7% 127.5% 136.0% 1.05 As it happens, the binomial tree C2in be 135.2% 1.10 100.3% 120.9% 126.5% used to value an LSS without becoming a forest. The key difference between an LSS and 99.8% 119,9% 125.4% 135,3% 1.15 an option on a CDO is that there are no decisions to be made by either party at any node 125.0% 136.0% 1.20 99.3% 119.0% in an LSS. 137.1% 98.8% 118.0% 124.8% 1.25 Denote the /th node at time T. by (/,/), Let S.. and E.. be the cumulative survival probability and expected tranche principal at node Note: Tlie tranches mature in five years. Results are hascd on llie three-parameter model calibrated lo the market data in Exhibit I. ((',/'). The value of 5. is given by Equation (4) and E. can be calculated from Ec]uations (2), (5), and (6). beginning of the transaction because that is all that the We first roll back through the tree calculating V.., seller ofthe protection is risking. As a result, the spread the value ofthe tranche to the protection buyer per dollar earned on this principal, assuming tbe tranche does not of principal at node {i,j) assuming no cancellation and experience defaults and the deal is not cancelled,, is LIBOR no limit on the liability ofthe protection seller.'^ Some plus 150 bps. times, r, correspond to payment dates and others do The model for 5 can be represented as a binomial not. Define 5 as follows. When T. is a payment date so tree. To construct the tree, the life ofthe model is divided that T. - Tj., 5. equals the accrual fraction t^ - f^_,. When into a number of short time intervals. Denote the time corr. is not a payment date, 5- ~ 0. Variables A.., B.,, and C.. responding to the end ofthe I'th interval by X. and let X,, = can be defined as follows: A. = S.E.., B = 0, and C = 0 0. During each time interval it is assumed that there is either zero or one jump in 5. This leads to a tree with the geometry shown in Exhibit 11. In this exhibit, H, is the size of thejthjump and M = M{T). The probability on the upper and lower branches emanating from a node at time T. are .^A.and 1-AA., respectively, where X. A(r.) and A. = tj+, r. The T. are chosen so that there are nodes
'.I

'

'J

at the final nodes, and at earlier nodes they are calculated by working backward through the tree using

22

I>r-NAM[c: MODELS OF PORTFOLIO CREDIT RISK: A SIMPLIFIED APPROACH

SUMMER 2008

C. =

To value the LSS it is necessary to roll back through the tree again. At node {i,J) we know the probability that a conipany has not defaulted, S... The probability of ti defaults by node (/,_/). is given by Equation (5) as h(u, N, 1 - SX The probability that termination has not been triggered at a node (ij) so that the deal is alive is therefore
(17)

where T* = 0.5(T + T _,). The P^,, are then calculated as

IJ

,, + B.)

As in the case of a CDO we assume without loss of generality that the principal is SI. Variables A'.-'^-^ and

EXHIBIT 11 Four-Step Binomial Tree for the Variable X, Which Is Minus the Log of the Cumulative Survival Probability

Time

Note: A/ 15 the value of X at time T. when there are no jiiwps; H. is the size of the Jth jtinipy. Tlie prohcihility on the upper iiiid lower Imituhes ematuiting from o node at time t. arc X^A^ and /-AA., respectively, where X.= X(Tf and A^ = r^, - r. The value of the lunitihitii't- siirt'ii'al pivluihihty. S, at any node is exp(-X}.

SUMMKR 2008

T H F jOURNAI. OF ilP-RtVATlVRS

23

Model with a^ - 0.12, d ^ 0.22. We considered leverage ^ levels between 5% {x = 0.05) and 20% (x = 0.2) and for trigger levels from 1 to 16 detaults. The results are shown in Exhibit 12. It should be noted that while the market spread for the 12% to 22% iTraxx tranche is 2 bps, the spread for this tranche in the calibrated model is 1.53 bps. All the results we present (tor forward contracts, options, and LSSs) are of course based on the calibrated model. In particular the spread for the LSS converges to the calibrated 1.53 bps. If our 1. The structure is still alive. sole objective had been to price the LSS considered here., 2. Cancellation was triggered at an earlier node. when carrying out the calibration we would have given 3. Cancellation is triggered at node (/, /). the squared error in the spread for the 12 to 22% tranche a high weight compared with that for other tranches. This The probability that the structure is still alive at node would have resulted in that tranche having a spread very (i,;') is w... In this case the backwards induction equations close to 2 bps in the calibrated model. for A'^'^ and C'r^'' are the same as those tor A., and C in (' U 'J 'J As the degree of leverage and the number of defaults Equation (17). required to trigger termination are increased, the The probability that the structure is not alive at node breakeven spread declines. Higher leverage means that (i, j) is 1 - w... We are then in situation 2 or situation 3. the buyer of protection gets a smaller payoff in the event Assume that situation 3 applies. The protection buyer must of default and so should be required to pay less for the promake a final payment equal to EE , where 8 is the time tection. When the number of defiuilts required to trigger since the previous payment date. There are two compotermination is increased, it is less likely that the deal will nents to C':'^'^. The seller of protection must pay -K., the be terminated prematurely and more likely that the seller present value of future net payments on a regular CDO of protection will benefit from the loss cap. Again, this tranche as well as the payment that would be due on a regreduces the amount the buyer of protection should be ular CDO tranche at time iAt. The latter is 1 E.. because required to pay. we know that the protection seller is not responsible for defaults prior to cancellation. Payoffs are capped at x so EXTENSIONS OF MODEL that C?:''"' equal the minimum of .v and K. + 1 E .. Assuming situation 1 or situation 3 applies leads to We now consider a number of ways in which the the following formulas for calculating Alr^'"" and Cjr'''' basic homogeneous constant recovery rate model can be extended. are defined analogously to A., and C so that sAb'''' is the present value of payments on the LSS from node {i,j) to the end ofthe life ofthe LSS and C'r'^^ as die present value of payoffs from node (/, /) to the end ofthe life ofthe transaction. (We do not need B'.'''' because the protection seller is never responsible for any defaults prior to cancellation in the LSS.) At node (i, /) we are in one ot three situations:
,^, ,

Stochastic Recovery Rate X i'(T.,,)/r(T.) + (1 - u'..)min[-K. + (1 - ..), x] Because ofthe way the values of Ah'^^ and C^^^ get overwritten as we work back through the tree, the values calculated at thefirstnode take into account the possibility of situation 2 applying at nodes (/,/). The breakeven spread is therefore C^'^^f^/A^'j^''. To illustrate the properties of LSS transactions we calculated the breakeven spread for a 5-year CDO iTraxx tranche on January 30, 2007, using the three-parameter Let the cumulative default probability, 1 S, be denoted as Q. A negative relationship between the recovery rate and the cumulative default rate can be incorporated into the model by assuming that the average recovery rate R applying to all the defaults that have occurred up to time t is a function only ofthe cumulative default probability, Q. One possibility is the relationship

aQ

24

DYNAMIC MODELS OF POIITFOLIO CREDIT RISK: A SIMPLIFIED APPROACH

SUMMEk 21)1)K

EXHIBIT 12 Breakeven Spread in a Leveraged Super Senior Transaction in Basis Points as a Function of the Number of Defaults Required to Trigger Termination for the iTraxx 12% to 22% Tranche on January 30, 2007 x = 0 05 - -' x = 0.10 X = 0.20

1.5

>

1.3

0.5

C 2 4 6 8 10 12

14

16

Number of Defaults Required to Trigger Termination

.\'ofc: Rc:!ti!is arc based on lhe ilircf-partimcicr model calibraivd lo lhe iimrkel data in Exhibir I. The maximum loss horne hy the seller of pmtcakm is x.

where a and R,, are positive constants and Q^- is the expected value of Q at the particular time being considered. Under this model, the marginal recovery rate when the cumulative default probability is Q is

Equation (2} rF(ff, ^ is replaced by py(n, / |y), R is replaced by R( J , 0 , and n^ and H^, are defined in terms of R so that they become dependent on J and f. Equation (6) becomes
I J) = | ; O ( , 11 J)W{n, 11 J)

This is greater than R,, when Q is lower than irs expected value and less than R,, when it is greater than its expected value. A constraint is that a must be chosen so that R^f'"'''^'^l^^ is always less than one. Defme R{j, t) as the average recovery rate between times 0 and ; wlien there have been J j u m p s in X. In

Apart from these changes the valuation of a CDO tranche is the same as before. Similar changes enable forward CDOs, European options on CDOs, and leveraged super seniors to be valued. In the case of leveraged super seniors, it is necessary to calculate the critical number of defaults at each node that lead to the loss threshold being exceeded.

SUMMER 2O(J8

THE JOURNAL OF DERIVATIVES

25

Heterogeneous Model The three-parameter model can be extended so that it becomes a heterogeneous model where each company has a different CDS spread.'-^ The model that has been presented is then assumed to represent the evolution of the cumulative default probability for a representative company in the portfoho. For any particular company in the portfolio, the jump size and jump intensity are assumed to be the same as that for the representative company. However, the deterministic drift, jU{r), is adjusted to match the CDS spread. The binomial model for determining the probability of n defaults by time f conditional on / jumps in Equation (5) must be replaced by an iterative procedure such as that in Andersen, Sidenius, and Basu [2003] or Hull and White [2004]. Once this has been done, the model structure and calculations for valuing a CDO are much the same as we have presented them, Valuing a forward CDO is similar to valuing a regular CDO. To value a leveraged super senior it is necessary to calculate at each node ofthe tree the probability the critical loss level has been exceeded. One appHcation ofthe iterative procedure just mentioned is therefore necessary for each node of tbe tree. For options on CDOs, calculations are more time consuming. It is necessary to calculate the probability of n - n^^ defaults between times t^^ and f, conditional on n^^ defaults by ^ time t^^ andj^jumps by this time. For this we require, for each ofthe ATcompanies, the probability of survival by time t^^ conditional on n^^ defaults and J^Jumps by time l^^. This is J^Y^^yy where /, is the unconditional probability of the company surviving to time f , } , is < the probability of out ofthe remaining N 1 companies defaulting, and J^ is the probability of ^ out of ^ N companies defaulting. A large number of applications of tbe iterative procedure is required.''* Modeling Two Portfolios Simultaneously The model can be extended so that iTraxx and CDX NA IG are modeled simultaneously. One way of doing this is to have three independent jump processes. The first process leads only to jumps in the cumulative hazard rate for iTraxx companies; the second jump process leads only to jumps in the cumulative hazard rate for CDX NA IG companies; the third jump process leads to jumps in the cumulative hazard rate for both iTraxx and CDX NA IG companies.

Bespoke Portfolios In practice it is often the case that derivatives dependent on bespoke portfolios have to be valued with dynamic models. The model we have presented must first be calibrated to iTraxx or CDX NA IG (or both). The drifts for the cumulative hazard rates of individual names comprising the bespoke portfolio can then be chosen to niatch their CDS spreads. CONCLUSION We have presented a simple one-factor model for the evolution of defaults on a portfolio. The model has two advantages over the Gaussian copula model. First, it is simpler and easier to implement. Second, it is a dynamic model that allows a wider range of products to be valued. The model is an alternative to the more complex dynamic models suggested by other researchers. To our knowledge this is the first article to use a dynamic credit model for pricing a variety of difFerent types of portfolio credit derivatives. The model has the attractive feature that it has many analytic properties and can be represented in the form of a binomial tree. The variable modeled on the tree is the cumulative survival probability for a representative company. The model is easy to use and appears to bave the property that fijture CDO spreads are approximately lognornial. We have shown how the model can be calibrated to market data. Our results indicate that in a risk-neutral world, there is a small chance that the default probability for a representative conipany during the hfe ofthe model will be very high. This is consistent with the results from static copula models. To fit market data it is necessary for jump sizes to become increasingly large. This is consistent with empirical data showing that default correlations are higher in recessionary periods. ENDNOTES We are grateful to Moody's Investors Service for providing financial support for this research. We are grateful to participants at the New York University Derivatives 2007 conference, the ICBI Global Derivatives Conference in May 2007, the Moody's/Copenhagen Business School conference in May 2007. and the University of Chicago conference on credit risk in October 2007. We would particuliidy like to thank Claudio Aibanese, Leif Andenen, Peter Carr, Rama Cont, Fierre

26

DYNAMK: MODELS OK PORTFOLIO CREDIT RISK: A SIMPLIFIED AI>PROACH

SUMMF.R 2008

CoUin-Dufresne, DarreU Duffie, Ruediger Frey, Michael Gordy, Igor Halperin, Jon Gregory, David Lando. Allan Mortensen, and Pliilipp Sfhonbuclier for useful comments and suggestions. We vvmild also like to thank the editor. Stcpiicn Figlewski, for extremely perceptive and helpful comments. 'This approach to separating default probabilities and default events is considered in Ehlers and Schonbuchcr |2(IO6] and is valid in our set up. -Our terminology reflects the viewpoint ofthe buyer of protection. We will refer to the payments made by tlie buyer of protection as "payments," and the payment in the event of default by the seller of protection as a "payoff." 'Processes of the form dX = fUilt + odz where dz is a Wiener process and (T> 0 are inappropriate because they allow A' to decrease. We could assume a process where O = 0 and // " follows a positive diffusion process. However, this would be ditticult to handle and would not calibrate well to market data. In our experience, large jumps in A'are necessary to fit market datLi. "The default intensity, A, can be allowed to depend on .V. Tlic model is then less analytically tractable, but can still be represented as a binomial tree. "Following the usual conventions, the quotes in Exhibit 1 ire the rate of payment in basis points per year to purchase protection from defaults in the indicated range. The exception is the 1) to y/i\ tranche where the quote is the up-ftont payment as a percentage ofthe notional that is paid in addition to .SiK) bps per year. ''A I O-day moving average provides a better indication of parameter values than the daily results because ofthe impact of noise in the quotes and missing quotes. ^Forward CDOs provide a simple application ofthe model and lead into the calculation of options on CDOs in the next section. As explained in Hull and White [2007] a dynamic model is not necessary to determine forward CDO spreads. ''We can also value options using the top-down version ofthe model. Only _ is then used in generating the exercise ^ decision. The results can be used to value options analytically in this case. Conditioning exercise on n^^ results in substantially lower option prices than conditioning on both it and 5. The differences vary by tranche and are largest for the mezzanine tranche. The average price reduction for the options in Exhibit Kis about 13%. ''For ease of exposition we assume if is constant. The methodology we present can easily be extended to the situation (often encountered in practice) where u' changes witb the passage of time. '"Our methodology can be extended to cases where this condition is not satisfied. "Similar to the development ofthe top-down version of the model in which only the numher of defaults is known, it is possible to develop a version ofthe model in which only S{t) is known. The resulting prices for European options on CDO

tranches are usually about 1% lower than when the exercise decision is based on both S{t) and the number of defaults to date. '-This could be calculated analytically, but since the tree is used for incorporating the uiipact ofthe cancellation it makes sense to use the tree for this as well. The results are slightly different from those previously reported because now defaults are no longer restricted to occur in the middle of an accrual period. ''Recall that we are working a filtration where, in addition to S((), we know the number of detiuilts by time f. We do not know the particular companies that have defaulted. Incorporating the latter into our model, or any other model, would be prohibitively time consuming. '"It may well be possible to speed up the calculations by developing a robust approximation to the probability that a particular company will survive by time f ^ conditional on u^ defaults.

REFERENCES
Albanese, C , O. Chen, A. Dalessandro, and A. Vidler. "Dynamic Credit Correlation Modeling." Working paper. Imperial College, 2005. Andersen, L., J. Sidenius, and S. Basu. "All Your Hedges in One Basket." Riik, 16 (November 2003), pp. 67-72. Ang, A., and J. Chen. "Asymmetric Correlations of Equity Portfolios." fotmial ofFinancia! Eamomics, 63 (2002), pp, 443-94. liaxter, M. "Levy Process Dynamic Modeling of Single Name Credits .md CDO Tranches." Working paper, Nomuni International, 2006. Beiinani, N. "The Forward Loss Model: A Dynamic Term Structure Approach for the I'ricing of Portfolio Credit Risk." Working paper, Royal Bank of Scotland, 2006. Brigo, D., A. F'allavicini, and R. Torresetti. "Calibration of CDO Tranches with the Dynamical Generalized-Poisson Loss Model." Risk. 20 (May 2007), pp. 70-75. Chapovsky, A., A. Rennie, and P.A.C. Tavares. "Stochastic Modeling for Structured Credit Exotics." Working paper, Merrill Lynch International, 20{J6. Das, S.R., D. Duffie, N. Kapadia, and L. Saita. "Common Failings: How Corporate Defaults are Correlated."7(>f(nw/ cif h'inatue., 62(2007), pp. 93-117. Das, S.R., L. Freed, C. Geng, and N, Kapadia. "Correlated
Default Kisk" Joiirtial of Eixed liicotnc, 16 (2006), pp. 7-32.

SUMMER 2008

THU JOURNAL OF I:)ERIVATIVES

27

Dutlic, D., and N. Garleanu. "Risk and the Valuation ofColHurd, T.R., and A. Kuznetsov. "Fast CDO Computations in AfFine lateralized Debt Obligations." Financial Analyst.': Joiiwat. 57 Qan- Markov Models." Working paper, McMaster University, 2005. uary/February 2001), pp. 41-59. Joshi, M., and A. Stacey. "Intensity Gamma." Risk, 19 (July Ehlers, P., and P. Schonbucher. "Background Filtrations and 2006), pp. 78-83. Canonical Loss Processes for Top-Down Models of Portfolio Credit Risk." Working paper, ETH Zurich, 2006. Li, D.X. "On Default Correlation: A Copula Approach ."Jen mu/ of Fixed Imome, 9 (2000), pp. 43-54. Brrais, E., K. Gieseke, and L. Goldberg. "Pricing Credit from the Top Down Using Aff me Point Processes." Working paper, LongstafF, H, and A. Rajan. "An Empirical Analysis ofthe Stanford Univenity, 2006. Pricing of Collateralized Debt Obligations." Jowma/ of Finance. Vol. 63, No. 2 (April 2008), pp. 529-563. Graziano, G., and L.C.G. Rogers. "A Dynamic Approach to the Modeling of Correlation Credit Derivatives Using Markov Merton, R.C. "On the Pricing of Corporate Debt: The Chains." Working paper. University of Cambridge, 2005. l^sk Structure of Interest Rates."_/oumi3/o/'Fmancp, 29 (1974). pp. 449-470. Gregory, J., andJ-P. Laurent. "Basket Default Swaps, CDOs, and Factor Copuias."J(iMrn(i/ of Risk.. 1 (2005). pp. H-23. Putyatin, V, D. Prieul, and S. Maslova. "A Markovian Approach to Modelling Correlated Defaults." Risk, 18 (May Heath, D , R. Jarrow, and A. Morton. "Bond Pricing and 2005), pp. 65-69. the Term Structure of Interest Rates: A New Methodology." Ecoiwmelrka, 60 (1992), pp. 77-105. Schonbucher, P. "Portfolio Losses and the Term Structure of Loss Transition Rates: A New Methodology for Pricing PortHull, J., and A. White. "Efficient Procedures for Valuing Eurofolio Credit Derivatives." Working paper, ETH Zurich, 2005. pean and American Path-Dependent Options." Journal ofDerwatives, 1 (1993), pp. 1-13. Servigny, A., and O. Renault. "Default Correlation: Empirical Evidence." Working paper. Standard and Poors, 2002. . "Valuation of a CDO and nth to Default CDS without Monte Carlo Simulation." yonmti/ of Derivatives, 12 (2004), Sidenius,J., V Piterbai^, and L. Andersen. "A New Framework pp. 8-23. for Dynamic Credit Portfolio Loss Modeling." Working paper. Bank of America, 2004. , "The Valuation of Credit Default Swap Options."' of Derivatives. 10 (2003), pp. 40-50. Vasicek, O. "Probability of Loss on a Loan Portfolio." Working paper, KMV, 1987. (Published m Risk in Dec. 2002, with the . "Valuing Credit Derivatives Using an Implied Copula title "Loan Portfolio Value.") Approach." JoHDiij/y/DcnVfl/iVps, 14 (2006), pp. 8-28. Walker, M. "Simultaneous Calibration to a Range of Portfolio . "Forward and European Options on CDO Tranches." Credit Derivatives with a Dynamic Discrete-Time Multi-Step Journal of Credit Risk, 3 (2007), pp. 63-73. Loss Model." Working paper, University of Toronto. Hull, J., M. Predescu, and A. White. "The Valuation of Correlation-dependent Credit Derivatives Using a Structural Model." Working paper. University of Toronto, 2005.

To order reprints of this article, please contact Dcwey Palmieri at (ipalmieri@iijoiirnals.com or 212-224-3675

28

DYNAMIC MOUF-US OF PORTFOLIO CREDIT RISK: A SIMPLIFIED APPKJLIACH

SUMMER 2(X)H